A deep clustering by multi-level feature fusion

Deep clustering extracts non-linear features through neural networks to improve the clustering performance. At present, deep clustering algorithms mostly only use single-level features for clustering, ignoring shallow features information. To address this issue, we propose a joint learning framework that combines features extraction, features fusion and clustering. Different levels of features are extracted through dual convolutional autoencoders and fused. Moreover, the clustering loss function jointly updates the dual network parameters and cluster centers. The experimental results show that the proposed network architecture fusing different levels of features effectively improves clustering results without increasing model complexity. Compared with traditional and deep clustering algorithms, the Clustering Accuracy (ACC) and the Normalized Mutual Information (NMI) metrics are significantly improved.